def main():
batch_size = 1
seq_len = 24
assert batch_size == 1
test = LPRIter(
path_imgrec='/opt/data/plate/color_rec/blue_vertices.rec',
path_imglist='/opt/data/plate/color_rec/vpr_blue_vertices.txt',
label_width=8,
# mean_img = '/opt/data/plate/rec/mean_plate.bin',
data_shape=(3, 94, 24),
batch_size=batch_size,
augment=False)
mod = load_module(
'/opt/models/mxnet/plate/plate_lprnet/deploy_lprnet_concat', 500,
test.provide_data)
batchnum = 0
hit = 0.
total = 0.
test_error_list = []
fsequence = open('/opt/incubator-mxnet/example/ctc/sequence.txt', 'w')
while True:
try:
batch = test.next()
input_data = batch.data[0].asnumpy()
batchnum += 1
mod.forward(batch)
output = mod.get_outputs()[0].asnumpy()
topk = ctc_decoder.decode(output)
label = CtcMetrics._remove_blank(batch.label[0].asnumpy()[0])
is_match = top_k_match(label, topk)
if is_match:
hit += 1
else:
decode_l = index_2_char(label)
decode_p = index_2_char(topk[0][0])
print_str = decode_l + '\t' + decode_p
print("label:%s\tpred:%s" % (decode_l, decode_p))
test_error_list.append(print_str)
total += 1.0
print(total)
except StopIteration:
break
ferror = open('/opt/incubator-mxnet/example/ctc/test_topk_error_list.txt',
'w')
for file_str in test_error_list:
ferror.writelines(file_str + '\n')
ferror.close()
print(hit / total, hit, total)
fsequence.close()
def infer(self, inputs, labels):
x, y, x_lens, y_lens = self.collate(inputs, labels)
out, probs = self.forward_imp(x, softmax=True)
probs = probs.cpu().detach().numpy()
x_lens = x_lens.cpu().detach().numpy()
res = [
decode(p[:l], beam_size=self.beam_size, blank=self.blank)[0]
for p, l in zip(probs, x_lens)
]
return res
def eval_model(self, inputs, labels, beam_size=1):
# Calculate loss and CER together
self.eval()
x, y, x_lens, y_lens = self.collate(inputs, labels)
out, probs = self.forward_imp(x, softmax=True)
loss = self.ctc_loss(out, y, x_lens, y_lens)
# Turn to cpu
probs = probs.cpu().detach().numpy()
# Use CTC decoder beam size to get prediction of characters
x_lens = x_lens.cpu().detach().numpy()
if beam_size == 1:
pred = [
max_decode(p[:l], blank=self.blank)
for p, l in zip(probs, x_lens)
]
else:
pred = [
decode(p[:l], beam_size=beam_size, blank=self.blank)[0]
for p, l in zip(probs, x_lens)
]
return loss.item(), pred